USERS_GUIDE.md 16.1 KB
Newer Older
Andreas Marek's avatar
Andreas Marek committed
1
## Users guide for the *ELPA* library ##
Andreas Marek's avatar
Andreas Marek committed
2

3
4
5
6
7
This document provides the guide for using the *ELPA* library with the new API (API version 20170403 or higher).
If you want to use the deprecated legacy API (we strongly recommend against this), please refer to the document
[USERS_GUIDE_DEPRECATED_LEGACY_API.md] (USERS_GUIDE_DEPRECATED_LEGACY_API.md).

If you need instructions on how to build *ELPA*, please look at [INSTALL.md] (INSTALL.md).
Andreas Marek's avatar
Andreas Marek committed
8
9
10
11
12

### Online and local documentation ###

Local documentation (via man pages) should be available (if *ELPA* has been installed with the documentation):

13
For example "man elpa2_print_kernels" should provide the documentation for the *ELPA* program, which prints all
Andreas Marek's avatar
Andreas Marek committed
14
the available kernels.
Andreas Marek's avatar
Andreas Marek committed
15

16
Also a [online doxygen documentation] (http://elpa.mpcdf.mpg.de/html/Documentation/ELPA-2018.11.001.rc1/html/index.html)
Andreas Marek's avatar
Andreas Marek committed
17
18
for each *ELPA* release is available.

Andreas Marek's avatar
Andreas Marek committed
19

20
### API of the *ELPA* library ###
Andreas Marek's avatar
Andreas Marek committed
21

22
23
With release 2017.05.001 of the *ELPA* library the interface has been rewritten substantially, in order to have a more generic 
interface and to avoid future interface changes.
Andreas Marek's avatar
Andreas Marek committed
24
25

For compatibility reasons the interface defined in the previous release 2016.11.001 is also still available
26
**IF AND ONLY IF** *ELPA* has been build with support of this legacy interface.
Andreas Marek's avatar
Andreas Marek committed
27

28
The legacy API defines all the functionality as it has been defined in *ELPA* release 2016.11.011. Note, however,
29
that all future features of *ELPA* will only be accessible via the new API defined in release 2017.05.001 or later.
Andreas Marek's avatar
Andreas Marek committed
30

31
32
33
34
35
36
37
38
39
40
41
42
43
As mentioned, we advise against it, but if you want to use the legacy API please look at the document 
[USERS_GUIDE_DEPRECATED_LEGACY_API.md] (USERS_GUIDE_DEPRECATED_LEGACY_API.md).

### Table of Contents: ###

- I)   General concept of the *ELPA* API
- II)  List of supported tunable parameters
- III) List of computational routines
- IV)  Using OpenMP threading
- V)   Influencing default values with environment variables
- VI)   Autotuning

## I) General concept of the *ELPA* API ##
Andreas Marek's avatar
Andreas Marek committed
44

45
Using *ELPA* just requires a few steps:
Andreas Marek's avatar
Andreas Marek committed
46

Andreas Marek's avatar
Andreas Marek committed
47
48
- include elpa headers "elpa/elpa.h" (C-Case) or use the Fortran module "use elpa"

Andreas Marek's avatar
Andreas Marek committed
49
- define a instance of the elpa type
Andreas Marek's avatar
Andreas Marek committed
50

Andreas Marek's avatar
Andreas Marek committed
51
- call elpa_init
Andreas Marek's avatar
Andreas Marek committed
52

Andreas Marek's avatar
Andreas Marek committed
53
- call elpa_allocate to allocate an instance of *ELPA*
Andreas Marek's avatar
Andreas Marek committed
54
55
56
57
58
59
60
61
62
  note that you can define (and configure individually) as many different instances
  for ELPA as you want, e.g. one for CPU only computations and for larger matrices on GPUs

- use ELPA-type function "set" to set matrix and MPI parameters

- call the ELPA-type function "setup"

- set or get all possible ELPA tunable options with ELPA-type functions get/set

Andreas Marek's avatar
Andreas Marek committed
63
64
- call ELPA-type function solve or others

Andreas Marek's avatar
Andreas Marek committed
65
- if the ELPA object is not needed any more call ELPA-type function destroy
Andreas Marek's avatar
Andreas Marek committed
66

Andreas Marek's avatar
Andreas Marek committed
67
- call elpa_uninit at the end of the program
Andreas Marek's avatar
Andreas Marek committed
68

69
70
71
72
73
74
75
76
To be more precise a basic call sequence for Fortran and C looks as follows:

Fortran synopsis

```Fortran
 use elpa
 class(elpa_t), pointer :: elpa
 integer :: success
Andreas Marek's avatar
Andreas Marek committed
77

78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
 if (elpa_init(20171201) /= ELPA_OK) then        ! put here the API version that you are using
    print *, "ELPA API version not supported"
    stop
  endif
  elpa => elpa_allocate()

  ! set parameters decribing the matrix and it's MPI distribution
  call elpa%set("na", na, success)                          ! size of the na x na matrix
  call elpa%set("nev", nev, success)                        ! number of eigenvectors that should be computed ( 1<= nev <= na)
  call elpa%set("local_nrows", na_rows, success)            ! number of local rows of the distributed matrix on this MPI task 
  call elpa%set("local_ncols", na_cols, success)            ! number of local columns of the distributed matrix on this MPI task
  call elpa%set("nblk", nblk, success)                      ! size of the BLACS block cyclic distribution
  call elpa%set("mpi_comm_parent", MPI_COMM_WORLD, success) ! the global MPI communicator
  call elpa%set("process_row", my_prow, success)            ! row coordinate of MPI process
  call elpa%set("process_col", my_pcol, success)            ! column coordinate of MPI process
Andreas Marek's avatar
Andreas Marek committed
93

94
  succes = elpa%setup()
95

96
97
98
99
  ! if desired, set any number of tunable run-time options
  ! look at the list of possible options as detailed later in
  ! USERS_GUIDE.md
  call e%set("solver", ELPA_SOLVER_2STAGE, success)
Andreas Marek's avatar
Andreas Marek committed
100

101
102
103
104
  ! set the AVX BLOCK2 kernel, otherwise ELPA_2STAGE_REAL_DEFAULT will
  ! be used
  call e%set("real_kernel", ELPA_2STAGE_REAL_AVX_BLOCK2, success)

105
106
107
108
  ! use method solve to solve the eigenvalue problem to obtain eigenvalues
  ! and eigenvectors
  ! other possible methods are desribed in USERS_GUIDE.md
  call e%eigenvectors(a, ev, z, success)
109

110
111
  ! cleanup
  call elpa_deallocate(e)
112

113
114
  call elpa_uninit()
```
Andreas Marek's avatar
Andreas Marek committed
115

116
117
118
C Synopsis:
```C
   #include <elpa/elpa.h>
119

120
121
   elpa_t handle;
   int error;
Andreas Marek's avatar
Andreas Marek committed
122

123
124
125
126
   if (elpa_init(20171201) != ELPA_OK) {                          // put here the API version that you are using
     fprintf(stderr, "Error: ELPA API version not supported");
     exit(1);
   }
Andreas Marek's avatar
Andreas Marek committed
127

128
   handle = elpa_allocate(&error);
Andreas Marek's avatar
Andreas Marek committed
129

130
131
132
133
134
135
136
137
138
   /* Set parameters the matrix and it's MPI distribution */
   elpa_set(handle, "na", na, &error);                                           // size of the na x na matrix
   elpa_set(handle, "nev", nev, &error);                                         // number of eigenvectors that should be computed ( 1<= nev <= na)
   elpa_set(handle, "local_nrows", na_rows, &error);                             // number of local rows of the distributed matrix on this MPI task 
   elpa_set(handle, "local_ncols", na_cols, &error);                             // number of local columns of the distributed matrix on this MPI task
   elpa_set(handle, "nblk", nblk, &error);                                       // size of the BLACS block cyclic distribution
   elpa_set(handle, "mpi_comm_parent", MPI_Comm_c2f(MPI_COMM_WORLD), &error);    // the global MPI communicator
   elpa_set(handle, "process_row", my_prow, &error);                             // row coordinate of MPI process
   elpa_set(handle, "process_col", my_pcol, &error);                             // column coordinate of MPI process
Andreas Marek's avatar
Andreas Marek committed
139

140
141
   /* Setup */
   elpa_setup(handle);
Andreas Marek's avatar
Andreas Marek committed
142

143
144
145
   /* if desired, set any number of tunable run-time options */
   /* look at the list of possible options as detailed later in
      USERS_GUIDE.md */
Andreas Marek's avatar
Andreas Marek committed
146

147
   elpa_set(handle, "solver", ELPA_SOLVER_2STAGE, &error);
148
149
150
151
  
   // set the AVX BLOCK2 kernel, otherwise ELPA_2STAGE_REAL_DEFAULT will
   // be used
   elpa_set(handle, "real_kernel", ELPA_2STAGE_REAL_AVX_BLOCK2, &error)
Andreas Marek's avatar
Andreas Marek committed
152

153
154
155
   /* use method solve to solve the eigenvalue problem */
   /* other possible methods are desribed in USERS_GUIDE.md */
   elpa_eigenvectors(handle, a, ev, z, &error);
Andreas Marek's avatar
Andreas Marek committed
156

157
158
159
160
   /* cleanup */
   elpa_deallocate(handle);
   elpa_uninit();
```
Andreas Marek's avatar
Andreas Marek committed
161

162
## II) List of supported tunable parameters ##
Andreas Marek's avatar
Andreas Marek committed
163

164
The following table gives a list of all supported parameters which can be used to tune (influence) the runtime behaviour of *ELPA* ([see here if you cannot read it in your editor] (https://gitlab.mpcdf.mpg.de/elpa/elpa/wikis/USERS_GUIDE))
Andreas Marek's avatar
Andreas Marek committed
165

166
167
168
169
170
171
172
173
174
175
176
| Parameter name | Short description     | default value               | possible values         | since API version | 
| :------------- |:--------------------- | :-------------------------- | :---------------------- | :---------------- | 
| solver         | use ELPA 1 stage <br>  or 2 stage solver | ELPA_SOLVER_1STAGE          | ELPA_SOLVER_1STAGE <br> ELPA_SOLVER_2STAGE      | 20170403          |
| gpu            | use GPU (if build <br> with GPU support)| 0                           | 0 or 1             | 20170403          | 
| real_kernel    | real kernel to be <br> used in ELPA 2 | ELPA_2STAGE_REAL_DEFAULT    | see output of <br> elpa2_print_kernels    | 20170403          |
| complex kernel | complex kernel to <br>  be used in ELPA 2 | ELPA_2STAGE_COMPLEX_DEFAULT | see output of <br>  elpa2_print_kernels     | 20170403          |
| omp_threads    | OpenMP threads used <br> (if build with OpenMP <br> support) | 1 | >1 | 20180525 |
| qr | Use QR decomposition in <br> ELPA 2 real | 0 | 0 or 1 |  20170403  |
| timings | Enable time <br> measurement | 1 | 0 or 1 |  20170403  |
| debug | give debug information | 0 | 0 or 1 | 20170403  |
       
177

178
## III) List of computational routines ##
179

180
The following compute routines are available in *ELPA*: Please have a look at the man pages or  [online doxygen documentation] (http://elpa.mpcdf.mpg.de/html/Documentation/ELPA-2018.11.001.rc1/html/index.html) for details.
181
182


183
184
185
186
187
188
189
190
191
| Name         | Purpose                                                                 | since API version |
| :----------- | :---------------------------------------------------------------------- | :---------------- |
| eigenvectors | solve std. eigenvalue problem <br> compute eigenvalues and eigenvectors | 20170403  |
| eigenvalues  | solve std. eigenvalue problem <br> compute eigenvalues only             | 20170403  |
| generalized_eigenvectors | solve generalized eigenvalule problem <br> compute eigenvalues and eigenvectors | 20180525 |
| generalized_eigenvalues  | solve generalized eigenvalule problem <br> compute eigenvalues only             | 20180525 |
| hermitian_multiply       | do (real) a^T x b <br> (complex) a^H x b                                        | 20170403 |
| cholesky                 | do cholesky factorisation                                                       | 20170403 |
| invert_triangular        | invert a upper triangular matrix                                                | 20170403 |
192
| solve_tridiagonal        | solve EVP for a tridiagonal matrix                                              | 20170403 |
193
194


195
## IV) Using OpenMP threading ##
196

197
198
199
If *ELPA* has been build with OpenMP threading support you can specify the number of OpenMP threads that *ELPA* will use internally.
Please note that it is **mandatory**  to set the number of threads to be used with the OMP_NUM_THREADS environment variable **and**
with the **set method** 
200

201
202
203
```Fortran
call e%set("omp_threads", 4, error)
```
204

205
**or the *ELPA* environment variable**
206

207
export ELPA_DEFAULT_omp_threads=4 (see Section V for an explanation of this variable).
208

209
Just setting the environment variable OMP_NUM_THREADS is **not** sufficient.
210

211
This is necessary to make the threading an autotunable option.
212

213
## V) Influencing default values with environment variables ##
214

215
216
For each tunable parameter mentioned in Section II, there exists a default value. This means, that if this parameter is **not explicitly** set by the user by the
*ELPA* set method, *ELPA* takes the default value for the parameter. E.g. if the user does not set a solver method, than *ELPA* will take the default "ELPA_SOLVER_1STAGE".
217

218
The user can change this default value by setting an enviroment variable to the desired value.
219

220
221
222
223
The name of this variable is always constructed in the following way:
```
ELPA_DEFAULT_tunable_parameter_name=value
```
224

225
, e.g. in case of the solver the user can
226

227
228
229
```
export ELPA_DEFAULT_solver=ELPA_SOLVER_2STAGE
```
230

231
in order to define the 2stage solver as the default.
232

233
234
235
236
237
238
239
**Important note**
The default valule is completly ignored, if the user has manually set a parameter-value pair with the *ELPA* set method!
Thus the above environemnt variable will **not** have an effect, if the user code contains a line
```Fortran
call e%set("solver",ELPA_SOLVER_1STAGE,error)
```
.
240

241
## VI) Using autotuning ##
242

243
244
Since API version 20171201 *ELPA* supports the autotuning of some "tunable" parameters (see Section II). The idea is that if *ELPA* is called multiple times (like typical in
self-consistent-iterations) some parameters can be tuned to an optimal value, which is hard to set for the user. Note, that not every parameter mentioned in Section II can actually be tuned with the autotuning. At the moment, only the parameters mentioned in the table below are affected by autotuning.
245

246
There are two ways, how the user can influence the autotuning steps:
247

248
249
250
1.) the user can set one of the following autotuning levels
- ELPA_AUTOTUNE_FAST
- ELPA_AUTOTUNE_MEDIUM
251

252
253
Each level defines a different set of tunable parameter. The autouning option will be extended by future releases of the *ELPA* library, at the moment the following
sets are supported: 
254

255
256
257
258
259
260
| AUTOTUNE LEVEL          | Parameters                                              |
| :---------------------- | :------------------------------------------------------ |
| ELPA_AUTOTUNE_FAST      | { solver, real_kernel, complex_kernel, omp_threads }    |
| ELPA_AUTOTUNE_MEDIUM    | all of abvoe + { gpu, partly gpu }                      |
| ELPA_AUTOTUNE_EXTENSIVE | all of above + { various blocking factors, stripewidth, |
|                         | intermediate_bandwidth }                                |
261

262
263
2.) the user can **remove** tunable parameters from the list of autotuning possibilites by explicetly setting this parameter,
e.g. if the user sets in his code 
264

265
266
267
268
```Fortran
call e%set("solver", ELPA_SOLVER_2STAGE, error)
```
**before** invoking the autotuning, then the solver is fixed and not considered anymore for autotuning. Thus the ELPA_SOLVER_1STAGE would be skipped and, consequently, all possible autotuning parameters, which depend on ELPA_SOLVER_1STAGE.
269

270
The user can invoke autotuning in the following way:
271
272


273
Fortran synopsis
274

275
276
277
278
279
280
281
```Fortran
 ! prepare elpa as you are used to (see Section I)
 ! only steps for autotuning are commentd
 use elpa
 class(elpa_t), pointer :: elpa
 class(elpa_autotune_t), pointer :: tune_state   ! create an autotuning pointer
 integer :: success
282

283
284
285
286
287
 if (elpa_init(20171201) /= ELPA_OK) then
    print *, "ELPA API version not supported"
    stop
  endif
  elpa => elpa_allocate()
288

289
290
291
292
293
294
295
296
297
  ! set parameters decribing the matrix and it's MPI distribution
  call elpa%set("na", na, success)
  call elpa%set("nev", nev, success))
  call elpa%set("local_nrows", na_rows, success)
  call elpa%set("local_ncols", na_cols, success)
  call elpa%set("nblk", nblk, success)
  call elpa%set("mpi_comm_parent", MPI_COMM_WORLD, success)
  call elpa%set("process_row", my_prow, success)
  call elpa%set("process_col", my_pcol, success)
298

299
  succes = elpa%setup()
300

301
  tune_state => e%autotune_setup(ELPA_AUTOTUNE_MEDIUM, ELPA_AUTOTUNE_DOMAIN_REAL, error)   ! prepare autotuning, set AUTOTUNE_LEVEL and the domain (real or complex)
302

303
304
305
  ! do the loop of subsequent ELPA calls which will be used to do the autotuning
  do i=1, scf_cycles
    unfinished = e%autotune_step(tune_state)   ! check whether autotuning is finished; If not do next step
306

307
308
309
    if (.not.(unfinished)) then
      print *,"autotuning finished at step ",i
    endif
310

311
    call e%eigenvectors(a, ev, z, error)       ! do the normal computation
312

313
  enddo
314

315
  call e%autotune_set_best(tune_state)         ! from now use the values found by autotuning
316

317
318
  call elpa_autotune_deallocate(tune_state)    ! cleanup autotuning object 
```
319

320
321
322
323
C Synopsis
```C
   /* prepare ELPA the usual way; only steps for autotuning are commented */
   #include <elpa/elpa.h>
324

325
326
327
   elpa_t handle;
   elpa_autotune_t autotune_handle;                               // handle for autotuning
   int error;
328

329
330
331
332
   if (elpa_init(20171201) != ELPA_OK) { 
     fprintf(stderr, "Error: ELPA API version not supported");
     exit(1);
   }
333

334
   handle = elpa_allocate(&error);
335

336
337
338
339
340
341
342
343
344
345
346
   /* Set parameters the matrix and it's MPI distribution */
   elpa_set(handle, "na", na, &error);
   elpa_set(handle, "nev", nev, &error);
   elpa_set(handle, "local_nrows", na_rows, &error);
   elpa_set(handle, "local_ncols", na_cols, &error);
   elpa_set(handle, "nblk", nblk, &error);
   elpa_set(handle, "mpi_comm_parent", MPI_Comm_c2f(MPI_COMM_WORLD), &error);
   elpa_set(handle, "process_row", my_prow, &error);
   elpa_set(handle, "process_col", my_pcol, &error);
   /* Setup */
   elpa_setup(handle);
347

348
   autotune_handle = elpa_autotune_setup(handle, ELPA_AUTOTUNE_FAST, ELPA_AUTOTUNE_DOMAIN_REAL, &error);   // create autotune object
349

350
351
   // repeatedl call ELPA, e.g. in an scf iteration
   for (i=0; i < scf_cycles; i++) {
352

353
     unfinished = elpa_autotune_step(handle, autotune_handle);      // check whether autotuning finished. If not do next step
354

355
356
357
     if (unfinished == 0) {
       printf("ELPA autotuning finished in the %d th scf step \n",i);
      }
358
359


360
361
362
363
364
365
366
      /* do the normal computation */
      elpa_eigenvectors(handle, a, ev, z, &error);
   }
   elpa_autotune_set_best(handle, autotune_handle);  // from now on use values used by autotuning
   elpa_autotune_deallocate(autotune_handle);        // cleanup autotuning
   
```
367

368
  
369
370
371